AI agent loses first battle in opening coffee shop: Gemini's chaotic discounts result in losses, GPT's excessive stinginess leads to raw material shortages According to Beating monitoring, AI evaluation agency Andon Labs has released actual test data of its AI agent Mona operating physical coffee shops. In the first two months, Mona ran on the Gemini 3.1 Pro model. During operation, the model has almost no concept of profit. It not only crazily overpurchases raw materials, but is also easily influenced by customer language, giving large discounts or even giving away goods for free, and even admitting a customer's claim of a 99% discount without verification. This resulted in the coffee shop spending approximately $15000 on supplier and equipment procurement, but with sales of only $9000 and a net operating loss of nearly $6000 (if fixed costs such as rent and wages are included, the total expenditure is as high as $38000). Subsequently, the team switched the model to GPT-5.5. The new model showed obvious anxiety in the face of losses and immediately stopped blindly ordering. But this also leads to another extreme: due to insufficient procurement, fresh raw materials are out of stock. As of June 25th, the supply rate of menu items has dropped to 77% and 10 dishes have been forced to be taken down. At the same time, GPT-5.5 exhibits strong anti inducement and anti jailbreak capabilities, rejecting all customers who request special prices or use social media promotion to exchange for free food. [Original link]